Federated Learning as a Service (FLaaS) enables large-scale, privacy-preserving training across heterogeneous edge devices but incurs significant deployment and communication overhead when distributing high-capacity models. This paper introduces the integration of server-side knowledge distillation into FLaaS, producing compact student models that are deployed for federated training and inference. Distillation is implemented as a configurable service within the FLaaS orchestration layer and operates offline prior to federated training, preserving existing client-side logic. Experiments on CIFAR-10 show that FLaaS+KD reduces model transmission size by \(88.8\%\) and deployment energy by nearly \(90\%\) , at the cost of an approximate \(21\%\) reduction in accuracy. These results demonstrate that centralized KD enables scalable, energy-efficient, and privacy-preserving deployment in resource-constrained federated and edge environments.